nvidia_device_plugin_enable
Description
The nvidia_device_plugin_enable parameter deploys the NVIDIA GPU Device Plugin DaemonSet to your AKS cluster. This plugin exposes nvidia.com/gpu as a schedulable Kubernetes resource, allowing workloads to request GPU access.
When using Azure GPU VM sizes (e.g. Standard_NC, Standard_ND, Standard_NV families), AKS automatically installs the NVIDIA kernel drivers on the nodes. This parameter adds the device plugin layer on top, which is required for Kubernetes to advertise and schedule GPU resources.
The plugin is pinned to GPU-capable nodes using the convox.io/gpu-vendor=nvidia node label, which Convox automatically applies when it detects a GPU VM size.
Default Value
The default value for nvidia_device_plugin_enable is false.
Use Cases
- Machine Learning: Enable GPU acceleration for model training and inference.
- Video Processing: Accelerate encoding/decoding workloads with NVIDIA hardware.
- Scientific Computing: Run CUDA-accelerated HPC applications.
- Deep Learning Inference: Deploy GPU-backed inference servers (e.g. Triton, TorchServe).
Setting Parameters
This parameter must be set at rack installation time:
$ convox rack install azure <name> nvidia_device_plugin_enable=true
Or updated after installation:
$ convox rack params set nvidia_device_plugin_enable=true -r rackName
Setting parameters... OK
Additional Information
GPU workloads in your convox.yml request GPUs via the gpu service field. The node pool running GPU VMs should use an NC*, ND*, or NV* VM size (via node_type or additional_node_groups_config). See also nvidia_device_time_slicing_replicas for sharing a single GPU across multiple workloads.
